T-Head, Alibaba's chip unit, unveiled the Zhenwu V900 at the Apsara Conference in Hangzhou on September 22. Chief executive Eddie Wu, also known as Wu Yongming, called it the most powerful artificial intelligence chip in China today. He said it delivers three times the computing performance of the Zhenwu M890 from May. Mass production and commercial release are scheduled for the first quarter of 2027. The May roadmap had put that date in the third quarter of 2027 (Tom's Hardware, 2026; TechNode, 2026).

Each V900 carries 216 gigabytes of memory and 1,200 gigabytes a second of chip-to-chip bandwidth. Alibaba says an in-house interconnect switch can join more than 1,000 cards as one system and scale a cluster to 500,000 cards. The chip is built for training and inference and supports the lower-precision formats FP8 and FP4. Alibaba did not publish a floating-point operations figure, a process node, a foundry, or a power envelope (Tom's Hardware, 2026; Hardware Busters, 2026).

The M890 had 144 gigabytes of memory and 800 gigabytes a second of interconnect. Those two published steps are 1.5 times. The three-times computing claim is Alibaba's number. Independent labs have not published a side-by-side test.

Alibaba sized the chip for a larger Qwen model and a larger cloud

Wu said Alibaba plans models with 5 trillion to 10 trillion parameters. The current flagship, Qwen3.8-Max, is about 2.4 trillion parameters. Tom's Hardware puts Moonshot's Kimi K3 near 2.8 trillion. Those next Qwen training runs are the workload Alibaba cites for the 500,000-card cluster (Tom's Hardware, 2026).

Wu also set a target of more than 20 gigawatts of data-center capacity for Alibaba Cloud worldwide by 2032. He said customer demand still outruns what the company can supply. Alibaba has committed more than $53 billion over three years to artificial intelligence infrastructure. Citigroup maps the 20 gigawatt target to about $160 billion of external cloud revenue around fiscal 2033. That figure is the bank's estimate, not an Alibaba forecast (Data Center Dynamics, 2026; Citi via Zhitong Finance, 2026).

By May, T-Head had shipped more than 560,000 chips in the Zhenwu family. Alibaba later put the external customer count above 650, across cars, finance, models, energy, and factories. A Panjiu supernode that uses the V900 is planned for the first quarter of 2027. The follow-on Zhenwu J900 and new Yitian server processors are listed for the third quarter of 2027 (TrendForce, 2026; Fudzilla, 2026).

Chinese labs are substituting parts. Other regions still buy the leading factories.

Huawei's HiSilicon unit is the other large Chinese accelerator house. Its Ascend line is used inside China, where Nvidia's newest parts cannot ship. Cambricon and Moore Threads design similar silicon. SMIC is the main domestic factory. Those names are in the June chip-stack guide on this site. Alibaba adds the cloud, the switch, the supernode, and the Qwen model on top of its own accelerator.

Export rules block the newest extreme ultraviolet lithography tools from ASML. Chinese groups use older deep ultraviolet tools more times on the same wafer, bind several larger dies into one accelerator with stacked packaging, and fund open instruction sets such as RISC-V. T-Head has put research money into RISC-V. Software around the V900 is Alibaba's SAIL tools, not Nvidia's CUDA. Yields can be lower and unit cost higher. The spend is meant to keep a supply path open if another license is denied.

Outside China the leading parts still come from TSMC and Samsung factories, ASML lithography machines, Synopsys and Cadence design software, and high-bandwidth memory from SK Hynix and Samsung. Nvidia's CUDA still holds most of the training software Western labs already wrote. I covered that concentration in July when Synopsys, Cadence, and Siemens EDA sat on almost every data-center chip. That list produces the fastest silicon. It also puts a lot of risk in a few plants and a few tool vendors.

China is also adding capacity on older nodes, roughly 28 nanometers to 90 nanometers, for cars, power grids, appliances, and factory boards, and on silicon carbide and gallium nitride for power electronics. Those parts are a separate market from a 500,000-card training cluster. Extra supply there can change the price of everyday boards even if frontier accelerators stay scarce.

More chip brands do not guarantee cheaper training cards in four or five years

A long list of accelerator vendors does not, by itself, cut the price of a training cluster in 2030 or 2031. Those clusters are limited by high-bandwidth memory, by the packaging that joins memory to the compute die, by power at the site, and by the people who keep a large job running. Those four items sit with a short list of suppliers.

A cheaper bill is more likely on older-node logic, where China is adding a lot of capacity for cars and industrial boards, and on inference cards that do not need the newest memory stack. Frontier training is the last place to budget a large drop in accelerator price on a five-year calendar. Power and packaging have to clear first.

Wu has told analysts that putting more T-Head chips into Alibaba halls is part of how cloud margin is supposed to rise. That is an Alibaba plan for its own rooms. It is not a forecast for the world price of Nvidia parts.

Tighter export rules would hit tools, memory, and a few factories first

If export rules tighten again, the rest of the industry slows at the same points. ASML's extreme ultraviolet machines have no second source at the leading edge. TSMC still makes most of the world's advanced logic. High-bandwidth memory still runs through a small set of Korean lines, with Chinese producers adding volume. Design software for chips that Western labs tape out still runs through Synopsys and Cadence. I wrote in August about YMTC trying to list in Shanghai while Washington kept changing how it labeled the company. Memory policy can move as fast as lithography policy.

The V900 announcement still leaves out the process node, the yield, and how existing model code will run on SAIL instead of CUDA. A cluster of 500,000 cards also has to be cooled and networked. Alibaba can own more of that stack than a merchant chip vendor. The cluster still has to stay up.

About Shashi Bellamkonda

Loading bio…

LinkedIn Profile

Get new posts by email:

Subscribe

Powered by